Medical images often contain artificial markers added by doctors, which can negatively affect the accuracy of AI-based diagnosis. To address this issue and recover the missing visual contents, inpainting techniques are highly needed. However, existing inpainting methods require manual mask input, limiting their application scenarios. In this paper, we introduce a novel blind inpainting method that automatically completes visual contents without specifying masks for target areas in an image. Our proposed model includes a mask-free reconstruction network and an object-aware discriminator. The reconstruction network consists of two branches that predict the corrupted regions with artificial markers and simultaneously recover the missing visual contents. The object-aware discriminator relies on the powerful recognition capabilities of the dense object detector to ensure that the markers of reconstructed images cannot be detected in any local regions. As a result, the reconstructed image can be close to the clean one as much as possible. Our proposed method is evaluated on different medical image datasets, covering multiple imaging modalities such as ultrasound (US), magnetic resonance imaging (MRI), and electron microscopy (EM), demonstrating that our method is effective and robust against various unknown missing region patterns.

翻译：医学图像常包含医生添加的人工标记，这些标记可能对基于人工智能的诊断准确性产生负面影响。为解决此问题并恢复缺失的视觉内容，亟需图像修复技术。然而，现有修复方法需手动输入掩膜，限制了其应用场景。本文提出一种新型盲修复方法，无需指定图像中目标区域的掩膜即可自动完成视觉内容恢复。所提模型包含无掩膜重建网络和目标感知判别器。重建网络由两个分支组成，分别预测带人工标记的受损区域并同步恢复缺失的视觉内容。目标感知判别器借助密集目标检测器的强大识别能力，确保重建图像的任何局部区域均无法检测到标记。最终，重建图像可尽可能接近干净图像。我们在涵盖超声（US）、磁共振成像（MRI）及电子显微镜（EM）等多种成像模态的不同医学图像数据集上评估了所提方法，结果表明该方法能有效且鲁棒地应对各类未知缺失区域模式。